7 Ways Buy-Side Firms Structure Expert-Network Call Prep in the Agent Era

The pre-call prep step at buy-side firms used to be a 20-minute skim of an expert's bio and a hand-typed question list. That workflow is being pulled apart. AI research agents, MCP-connected transcript archives, and structured-financials tooling have turned prep into a stack of discrete, sequenced tasks, each with its own vendor pattern and its own compliance surface.

This guide maps the seven prep workflows a well-run research team now runs before an expert picks up the phone. It is written for analysts, PMs, and heads of research trying to decide which of these steps to formalize, and which to leave to the individual analyst.

1. Transcript pre-read via MCP retrieval

The first workflow is also the newest. Guidepoint's MCP server exposes more than 100,000 transcripts to Claude and Perplexity clients, so an analyst can ask their own agent to surface every prior call a given expert has done inside the network before scheduling a new one. That inverts the old order, where the analyst booked the call first and read the archive during the wait.

The archive-first pattern is not unique to one network. AlphaSense, following its Tegus acquisition, has been building toward a similar posture, with its expert transcript library retrievable from inside analyst-facing workflows. The practical effect is that the pre-call prep now starts with the question what does this expert already know that we already have access to, which materially reshapes the question list.

The risk INFLXD flags: agent-side retrieval only works if the citation layer is trustworthy. An agent that summarizes without anchoring to a specific transcript timestamp is a research liability, not a research asset.

2. Expert-vetting brief

The second workflow is bio-and-affiliation vetting. EN bio pages give tenure, title, and disclosed employers; the analyst's job is to confirm those against LinkedIn, Bloomberg, and, where relevant, regulatory filings. This is the step where AI copilots are least useful, because the failure mode is undisclosed affiliation rather than unstructured information.

The Action on Armed Violence finding that UK media outlets omitted defence-industry ties for roughly 60% of cited retired military officers is a useful reference point for why manual affiliation checks still matter, even outside financial contexts. A buy-side analyst preparing a call on, say, a defence prime cannot rely on a bio page to surface consulting arrangements the expert never disclosed.

The emerging practice: treat the bio page as the input, not the output. The output is a one-page vetting brief written by the analyst or a junior, listing every affiliation the analyst could verify and every one the EN attested to but the analyst could not.

3. Compliance gating

The third workflow sits between vetting and scheduling. It is not an analyst-side step; it is the EN's attestation flow. Networks including Guidepoint, GLG, and Third Bridge run MNPI screens, employer-restriction lists, and cooling-off period checks before a call is booked, and the analyst's role is to submit the correct restriction list and topic scope on the front end.

What has changed in the agent era is that compliance gating now has to survive the retrieval layer as well. If an agent can pull a prior transcript from the EN's archive, the EN's transcript-level redactions and restrictions have to travel with the retrieval, or the compliance perimeter leaks. This is one of the reasons MCP-style deployments have been slower to reach hedge funds than the underlying capability would suggest.

The honest read: this step is the least visible and the most consequential. A firm that lets agentic prep bypass EN compliance attestation has created a governance problem, not a productivity gain.

4. Question-set generation from thesis docs

The fourth workflow is where the AI-agent vendors are most visible. Rogo markets an agent that ingests IC memos and generates ranked expert questions; Hebbia's platform pushes a similar pattern from unstructured document sets; and Bridgetown Research has positioned its offering explicitly for PE diligence, backed by a USD 19M Series A.

The workflow looks the same across vendors: point the agent at the thesis, the target company's filings, and any prior transcripts, and ask for the questions the thesis needs answered but cannot answer from public data. In practice, the ranked list is a starting draft. Analysts INFLXD has spoken to describe re-ordering, cutting, and rewriting between 40% and 70% of what the agent produces, which is roughly the same edit ratio as a junior-analyst draft.

The strategic question for a head of research: is this workflow an analyst tool or a junior-replacement tool? The current answer, honestly, is the former. The agent is faster than a junior at producing a first draft; it is not yet better.

5. Prior-transcript diffing

The fifth workflow builds on the first. Once an analyst has pulled two or three prior transcripts on the same topic, the useful agent task is not summarization but diffing: given what these three experts have already said, what is still unknown, what is contradicted across the set, and where the consensus is thin.

This is where citation-anchored retrieval earns its keep. A diff that cannot be traced to a specific speaker and timestamp is a hallucination risk, and a diff that surfaces a contradiction between two experts is only useful if the analyst can click through to the exact exchange in each transcript. INFLXD has covered the transcript-citation standard emerging around this pattern; the short version is that timestamp-anchored retrieval is the minimum bar for agent-generated diffs to be safe to act on.

What a good diff prompt asks

• Which claims appear in two or more transcripts and agree?
• Which claims appear in two or more transcripts and conflict?
• Which claims appear only once, and is the source expert the most credentialed on that specific point?

6. Structured-data grounding

The sixth workflow brings quantitative data into the prep doc. Expert calls are qualitative by design, but the analyst's ability to pressure-test an expert's numbers live depends on having the company's KPIs, segment splits, and filings loaded and structured before the call starts. Daloopa and similar vendors sit in this slot, converting filings into structured financials that an agent can query without hallucinating.

Daloopa's own benchmark work suggests that grounding a retrieval-augmented generation flow in structured financials lifted accuracy by 71 percentage points against an unstructured baseline. The number is vendor-published and should be read as such, but the direction is consistent with what analysts report: an agent asked to compare an expert's claimed segment margin against the company's disclosed segment margin only works if the disclosed number is in a table, not scraped from a PDF footnote.

The practical output is a prep doc that puts the expert's likely claims next to the company's actual numbers, so the analyst can interrupt the call the moment the two diverge.

7. Post-call capture design

The seventh workflow is the one most often skipped, which is why INFLXD is listing it explicitly. Before the call happens, the analyst (or the research operations team) has to decide how the call will be recorded, transcribed, tagged, and rerouted into the firm's research library. That includes the transcription vendor (Aiera, AlphaSense, or an in-house pipeline), the tagging schema, the retention clock the EN enforces, and the access controls on the resulting transcript.

The reason to design this upfront rather than after the call: retention settings and access controls are contractual with the EN, and they interact with the compliance gating in workflow three. A transcript that lives in the firm's research library under one retention policy while the EN's contract specifies another is a compliance problem that surfaces months later, usually during an audit.

The emerging practice at larger buy-side firms is to codify capture design at the coverage-team level, not the individual-call level, so every analyst covering a given sector uses the same tagging schema and the same retention defaults.